Accurate segmentation of the prostate in computed tomography (CT) images is very important in image-guided radiotherapy. In the current study, an automatic framework is proposed for prostate segmentation in CT images: first, we propose a novel image feature extraction method, namely, variant scale patch, which can provide rich image information in a low dimensional feature space; second, we take the general idea of sparse representation and design a new segmentation criterion called local independent projection (LIP); third, we use an online update strategy to construct a dictionary to utilize the latest image information. Furthermore, in the proposed LIP, we emphasize locality rather than sparsity, and use local anchor embedding to solve the dictionary coefficients. The proposed method is evaluated based on 201 3D images of 12 patients. Results show that the proposed method is robust in segmenting prostates in CT images.